
Static S-Boot/LK probe for Samsung Exynos 1380 (s5e8835): SVE-2026-3546 / CVE-2026-21155 resolver patch, four-device coverage, and safety notes.
Static resolver probe for Samsung Exynos 1380 (s5e8835) S-Boot/LK — SVE-2026-3546 / CVE-2026-21155.
STATIC ONLY. This repository contains no USB code, no flashing code, no fuse code, and no device-write path. Nothing here talks to a phone. The tools read a local
sboot.binand print addresses. That is the whole attack surface.
The public SVE-2026-3546 proof of concept targets s5e8825 only. It resolves the s5e8825 LK image, builds a loader for its memory layout, and ends in a one-way Secure Boot KEY2 fuse. Owners of the newer Exynos 1380 / s5e8835 devices — Galaxy Tab S9 FE, Galaxy A26 5G, Galaxy A35 5G, Galaxy A54 5G — had no documented answer to the first question: does the resolver even understand this LK?
This repository answers that question statically. It documents the s5e8835 LK
layout, the two resolver changes needed to make parallel() and input_map()
recognize it, and the exact heuristics that still fail. The result was validated
against four independent sboot.bin images spanning 2025-08 to 2026-10.
No hardware was used. No device was unlocked. No fuse was burned.
sboot.bin by scanning for the BSS
descriptor triple, then validates the candidate with resolve.image_map().compressed_download_init instruction window and the MOVZ/MOVK
constants that input_map() sees, so a failing build can be diagnosed without
a disassembler.resolve.py patch needed for the s5e8835 slot-register layout
while preserving the s5e8825 behavior.resolve.py.All four images are sboot.bin extracted from official BL packages. The table is
the complete dataset behind the findings.
| Device | Model | Build | SPL | sboot.bin sha256 | LK slice | Base | BSS |
|---|---|---|---|---|---|---|---|
| Galaxy Tab S9 FE (Wi-Fi) | SM-X510 | X510XXSGEZI3 | 2026-09-05 | a8abc835e5b7337fd6ed6009abe74bde18882d8807841ae6d48e62eebf44f267 | 0xf1000 | 0xf4800000 | 0xf4c3d000..0xf507a908 |
| Galaxy A26 5G | SM-A266B | A266BXXS4AYGA | 2025-08-01 | 79f169c9ecc03f604add536c5d3eb3c1a097ecabb7d4c3986ad20387bcd481c4 | 0xf1000 | 0xf4800000 | 0xf4c3c000..0xf5105df8 |
| Galaxy A35 5G | SM-A356E | A356EXXSBDZH1 | Oct 2026 | 797fcba92b90e6ceadd55ffa6adaea64793c4a378c0dca26efc8ce9f89e37484 | 0xf1000 | 0xf4800000 | 0xf4c3c000..0xf5079908 |
| Galaxy A54 5G | SM-A546E | A546EXXUNGZJ5 | Oct 2026 | cb386232b2b32fc55fb3af7c676e54fb4fb43080e4d2adb1391edd8fdf5d9dc9 | 0xf1000 | 0xf4800000 | 0xf4c44000..0xf5087270 |
All four: end = 0xf6000000.
Download-mode buffers are identical across the family:
PAR_BASE = 0xce800000 PAR_SIZE = 0x7e00000 PAR_SLOT = 0x1e00000
IN_BASE = 0x890010000 IN_SIZE = 0x8000000 IN_SLOT = 0x2000000
The BSS range differs on every build. That is expected: it is linker output, not a
protocol constant. image_map() recovers it from the descriptor at the head of the
LK slice, so it must never be hardcoded.
Every image places the LK sub-image at file offset 0xf1000 and maps it at
0xf4800000, ending at 0xf6000000. The BSS descriptor sits at
0xf1030 (file offset), immediately inside the slice.
The whole sboot.bin is 0x800000 bytes. The first 0xf1000 bytes are EPBL and
other pre-LK blobs; the s5e8835 string at file offset 0x5e43 lives there, not
in LK. Additional marker copies appear inside LK on some builds:
| Device | s5e8835 marker file offsets |
|---|---|
| SM-X510 | 0x5e43 |
| SM-A266B | 0x5e43, 0x26e42b, 0x4f125f |
| SM-A356E | 0x5e43, 0x26ee28, 0x4f125f |
| SM-A546E | 0x5e43, 0x276bf0, 0x4f125f |
Both parallel_download_init and compressed_download_init materialize the same
buffer geometry as the s5e8825 ground truth. On the X510 the functions resolve at:
parallel_download_init @ 0xf4863d64
compressed_download_init @ 0xf4863fb0
The patch changes only how the slot register is recognized, not the buffer values. After the patch both rows return the exact s5e8825 values on every s5e8835 build tested.
w9 to w10The upstream resolver was written against s5e8825, where the per-slot size is
loaded into w9:
mov w9, #0x1e00000 ; PAR_SLOT (or #0x2000000 for IN_SLOT)
On s5e8835 LK, the same constant is loaded into w10, while w9 carries the
0x40 stride/count value:
mov w10, #0x2000000 ; slot size
mov w9, #0x40 ; slot count / stride constant
The upstream predicates slot[2:] != (9, 0, 0x52800000) therefore reject the real
slot on s5e8835 and the resolvers return an empty candidate list. The patch accepts
w9 or w10.
Simply accepting w10 everywhere is not safe. The two SoCs schedule the base,
total-size, and slot-size MOVZs in a different order:
w9): the 64-bit base pair is emitted before the size loads.
Upstream's base_offset >= min(total[0], slot[0]) rejection is part of the
original filter.w10): the base pair can be emitted after the slot load, so the
same rejection discards the valid candidate.The final patch applies the ordering rejection only when the slot register is
w9. That is the whole reason the A53 regression stays green.
An intermediate revision accepted w10 without the conditional. On the s5e8825
A53 image that admitted a spurious candidate and the regression line no longer
reproduced. The final slot[2] == 9 guard restores the ground truth:
base=0xf8800000 in=0x890010000 pack=0x894110000 code=0x896010100 out=0xc5000000 write=0xf8f03de8 cb=0xf8f04008 walk=0x33f03de8
Any future resolver change must reproduce that line before it is considered portable.
blk_bread and blk_bwrite resolve by symbol XREF and non-PAC prologue on two
builds:
| Device | blk_bread | blk_bwrite |
|---|---|---|
| SM-X510 | 0xf4846b58 | 0xf4846d54 |
| SM-A266B | 0xf4845fb8 | 0xf48461b4 |
Both wrappers use the classic non-PAC prologue:
fd7b bda9 f657 01a9 f44f 02a9 fd03 0091
A35/A54 wrapper addresses were not captured in the tester report. The rows are expected to pass there as well, but this repository does not claim an address it did not observe.
On the X510 and A26 after the patch, the nine-row probe is 4 PASS / 5 FAIL:
| Row | Failure signature |
|---|---|
compressed output block | RuntimeError: compressed output: [] |
Odin setter/reply | RuntimeError: Odin setter: [] |
screen printf/cursor | IndexError: list index out of range |
em_crypto_sha256 | RuntimeError: SHA-256 XREF not found |
blk shift | RuntimeError: block shift |
These are heuristic-shape failures, not evidence that the functions are absent: